All Languages Matter: Opening Swahili-to-Sukuma on the Edge

In 2024 we built Kamusi, a way to learn, look up, and explore Swahili with AI nearby. Even then, the point was never Swahili alone. We believe all languages matter, and that intelligence should live in the languages people actually speak and write, not only the ones that already crowd the internet.

Sukuma is spoken by millions. It should not wait at the edge of the map. So today we open the next chapter: two Swahili → Sukuma translators, shared as a research preview. They are small on purpose, for phones, everyday machines, and places where big cloud GPUs are scarce and the network is thin.

Type Swahili, get Sukuma (Kisukuma) back. Give it a try.

Why two versions?

Not every device has the same room. We trained two sizes on the same kind of literary Swahili–Sukuma text:

• Quality when you can afford a stronger model. On Hugging Face
• Lite when the device is tighter. On Hugging Face

What to expect

This is still a research preview. The path is Swahili → Sukuma only. English will not work as input: move to Swahili first, then translate. The tone leans literary, so casual chat, legal, or medical wording may sound odd. Use the box above as a demo, and the open models as a starting point for anyone who wants Sukuma in the loop.

For builders

We trained on about 31,000 Swahili–Sukuma pairs, keeping train and test apart by document. On our check set, quality lands around chrF2 42.6 and lite around chrF2 34.8. If you load the weights yourself, force the Sukuma target tag (suk_Latn) with Swahili as source (swh_Latn). The try box already does that. More detail lives in the technical report below.

Collaboration

Interested in extending this work, evaluating on new data, or building with Sukuma in your community? Write to us at hi@nileagi.com.

Weights are under CC BY-NC-SA 4.0. Commercial use needs a written agreement with NileAGI.